SD3BlocksMultiply

Same category as its sibling SD3Multiply - _for_testing/attention_experiments, straight from the author's own node listing. This one takes a different cut at the same idea: instead of scaling specific weight types uniformly across the whole model (attention, MLP, modulation), SD3BlocksMultiply gives you one multiplier per transformer block, whole. Twenty-four of them, b0 through b23.

What "block" means here

SD3's MMDiT architecture is a stack of transformer blocks, each processing a text-conditioning stream and an image-latent stream in parallel and letting them cross-inform each other. SD3BlocksMultiply scales every weight inside a given block index by the same number - turn b0 down and you're damping the entire first block's contribution to the pass, wholesale. This is a coarser tool than SD3Multiply, which instead lets you isolate attention versus MLP versus modulation across all blocks at once. Where SD3Multiply answers "which component matters," SD3BlocksMultiply answers "which depth in the network matters" - early blocks generally shape rougher, more global structure, later blocks tend toward finer detail, similar to how block position matters in other diffusion U-Nets and transformers, though SD3's specific per-block behavior isn't something either the README or general documentation breaks down in detail - this is exploratory territory, which is exactly why the author filed it under "testing."

The fields

b0 through b23, each a float, default 1 (unchanged), range 0–10. Twenty-four sliders is a lot to reason about at once - realistically you'd zero out or reduce one or two at a time and observe, not tune all of them together. Input: model, a loaded SD3 checkpoint. Output: MODEL.

Installing SP-Nodes

Manager: search "SP-Nodes," install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/bananasss00/ComfyUI-SP-Nodes

Restart once cloned. No model download needed for this node - it operates on weights already loaded elsewhere in the graph.

What to expect

Treat this the same way you'd treat any node the author explicitly filed under "testing": it's for probing and understanding a model's internals, not for wiring into a workflow you actually rely on. Setting a block's multiplier to 0 effectively removes that block's contribution entirely, which is a legitimate way to sanity-check "does this specific layer even matter for the output I care about" - but don't expect the result to still look like a normal generation. Expect visible breakage - smeared structure if you zero an early block, texture/detail loss if you zero a late one - proportional to how far from 1 you push it and which block you hit.

Worth keeping in mind going in: SD3's own standing took a real hit at launch (widely reported anatomy problems) and it was pulled entirely from CivitAI in 2025 over a license change, so most current advice steers people toward Chroma or newer models for the things SD3 used to cover. That doesn't make this node useless - if you're specifically trying to understand SD3's architecture, or you're one of the people still running the narrow set of things SD3.5 Large genuinely does well, block-level ablation like this is a legitimate way to learn what's happening inside it. Just don't go in expecting it to fix or improve a generation the way a merge or a LoRA would; it's a magnifying glass, not a tool.